Wafer test offset detection method, system, and storage medium

Abstract

The present application belongs to the technical field of semiconductor manufacturing, and provides a wafer test offset detection method, system and storage medium, which comprises: collecting key vector data in a wafer test process, the key vector data comprising a thermal load vector, a system state vector and a real-time observation vector, collating to obtain a multi-modal data set, establishing a hybrid prediction model combining a physical simulation model and a residual correction model, predicting an actual thermal displacement prediction vector based on the multi-modal data set using the hybrid prediction model, judging whether there is a thermal displacement mismatch risk, and if so, triggering offset root cause identification, establishing a historical knowledge base, and if the offset root cause identification is triggered, identifying the thermal displacement mismatch root cause of the wafer test based on the multi-modal data set and the historical knowledge base, generating a sequential root cause list, and sequentially generating a mismatch compensation strategy for the thermal displacement mismatch root cause according to the obtained sequential root cause list to compensate for the thermal displacement mismatch risk.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, specifically a wafer testing offset detection method, system, and storage medium. Background Technology

[0002] In the semiconductor manufacturing process, wafer testing is a key step in ensuring chip yield and performance. It uses probe cards to contact test points on the wafer to accurately measure the chip's electrical parameters. As chip manufacturing processes continue to break through to the nanometer level and wafer integration continues to increase, the alignment accuracy requirements between the probes and wafer test points during testing have reached the micrometer or even sub-micrometer level. However, the wafer testing environment is complex and variable, and misalignment problems occur frequently, becoming a core bottleneck restricting testing efficiency and accuracy.

[0003] The offset problem was discovered due to multiple anomalies in actual production: on the one hand, test data showed sudden dispersion, the parameter consistency of wafers in the same batch decreased significantly, and some chips were misjudged as defective products; on the other hand, the wear rate of probe cards accelerated abnormally, and even scratches appeared on the wafer surface, resulting in high production cost losses. Through retrospective analysis of the anomalies, it was found that the offset mainly originated from physical displacement caused by thermal effects. During the testing process, the wafer and probe card are affected by factors such as ambient temperature fluctuations, changes in power consumption of test items, and equipment self-heating, resulting in small but cumulative thermal expansion or contraction, leading to spatial mismatch between the probe and the test point.

[0004] Existing technologies have significant limitations in addressing wafer testing offset issues. Specifically, these limitations manifest in several ways: data acquisition is limited to a single dimension, key influencing factors are not fully captured, leading to a one-sided assessment of the offset's causes and making it difficult to construct a complete causal chain. Predictive models lack accuracy and struggle to balance physical interpretability with adaptability to complex scenarios. Existing prediction methods fall into two categories: one is based on a purely physical model, which can explain the mechanism of thermal displacement but cannot cover nonlinear factors such as probe wear and airflow disturbances, resulting in significant discrepancies between theoretical calculations and actual displacements; the other is based on a purely data-driven model, which can fit complex patterns but lacks physical support, resulting in poor interpretability of prediction results and insufficient generalization ability under extreme conditions, making it prone to misjudgments. Furthermore, once an offset occurs, existing technologies often rely on manual investigation by engineers, making it difficult to quickly pinpoint the root cause, and compensation strategies lack specificity and dynamic adjustment capabilities.

[0005] To address the above problems, this invention proposes a wafer test offset detection method, system, and storage medium. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve the technical problem is: a wafer testing offset detection method, comprising:

[0008] Key vector data is collected during wafer testing. The key vector data includes thermal load vector, system state vector and real-time observation vector, which are then processed to obtain a multimodal dataset.

[0009] A hybrid prediction model combining a physical simulation model and a residual correction model is established. Based on a multimodal dataset, the hybrid prediction model is used to predict the actual thermal displacement prediction vector. It is then used to determine whether there is a risk of thermal displacement mismatch. If so, the root cause identification of the offset is triggered.

[0010] Establish a historical knowledge base. If offset root cause identification is triggered, identify the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generate a sequential root cause list.

[0011] Based on the obtained sequential root cause list, mismatch compensation strategies are generated sequentially for thermal displacement mismatch root causes to compensate for the risk of thermal displacement mismatch.

[0012] The physical simulation model includes a temperature field solution sub-model and a structural displacement calculation sub-model.

[0013] The temperature field solution sub-model and the structural displacement calculation sub-model include:

[0014] In the temperature field solution sub-model, the thermal load vector in the multimodal dataset is obtained as input. The thermal load vector includes the ambient temperature sequence and the power consumption spectrum of the test item. The boundary conditions of the heat conduction equation are determined according to the ambient temperature sequence. The internal heat source intensity is calculated by combining the power consumption spectrum of the test item and the volume of the heat-generating area. Based on the internal heat source intensity, the temperature field of the wafer and probe card that changes with time and space is obtained by solving the heat conduction equation and output.

[0015] In the structural displacement calculation sub-model, the temperature field output by the temperature field solution sub-model is used as input. Based on Hooke's law, the theoretical thermal displacements in the x and y directions caused by thermal expansion are calculated and integrated to obtain the theoretical thermal displacement vector.

[0016] The residual correction model includes:

[0017] A multimodal dataset is obtained. All key vector data in the multimodal dataset are Z-score normalized and fused using an attention mechanism fusion algorithm to obtain a multimodal feature matrix. The multimodal feature matrix and the theoretical thermal displacement vector are used as inputs to the residual correction model. The network structure of the residual correction model includes three spatiotemporal convolutional layers and one fully connected output layer, which outputs the predicted residual vector.

[0018] The method for obtaining the actual thermal displacement prediction vector is as follows:

[0019] Set a historical time period, obtain the theoretical thermal displacement vector within the historical time period and the corresponding actual thermal displacement vector obtained through detection and acquisition, calculate the corresponding actual residual vector, integrate it with the multimodal feature matrix to obtain the correction training set, train the residual correction model, use the trained residual correction model to perform deduction to obtain the predicted residual vector, and superimpose the theoretical displacement vector calculated by the physical simulation model with the predicted residual vector predicted by the residual correction model to obtain the actual thermal displacement prediction vector.

[0020] The historical knowledge base is established in the following way:

[0021] Establish a historical knowledge base to store feature-cause mapping pairs of past offset events, set historical time periods, and past offset events represent events with thermal displacement mismatch risk within the historical time period. The feature-cause mapping pairs include the multimodal dataset of past offset events, the corresponding actual thermal displacement vectors collected, and the root causes of thermal displacement mismatch.

[0022] The method for identifying the root cause of thermal displacement mismatch is as follows:

[0023] If offset root cause identification is triggered, a multimodal dataset is obtained, and the cosine similarity between the multimodal dataset and the multimodal dataset of past offset events stored in the historical knowledge base is calculated. If the cosine similarity is greater than the preset similarity standard, the thermal displacement mismatch root cause of the past offset event is included in the candidate root cause list as the identified thermal displacement mismatch root cause.

[0024] The sequential root cause list is generated as follows:

[0025] The frequency of occurrence of thermal displacement mismatch root causes in the candidate root cause list is obtained and processed with the total number of occurrences of thermal displacement mismatch root causes in the candidate root cause list to obtain the frequency of occurrence of thermal displacement mismatch root causes. All actual thermal displacement vectors corresponding to thermal displacement mismatch root causes in the historical knowledge base are obtained and processed to obtain the normalized thermal displacement vector of each type of thermal displacement mismatch root cause. The analytic hierarchy process (AHP) is used to assign weights to the frequency of occurrence of thermal displacement mismatch root causes and the normalized thermal displacement vector and perform weighted fusion processing to obtain the priority coefficient of thermal displacement mismatch root causes. The thermal displacement mismatch root causes in the candidate root cause list are sorted in descending order based on the priority coefficient to generate the ordered root cause list.

[0026] The method for compensating for the risk of thermal displacement mismatch is as follows:

[0027] A compensation strategy library is established, which stores the optimal mismatch compensation strategy corresponding to various thermal displacement mismatch root causes. The optimal mismatch compensation strategy is selected by obtaining the mismatch compensation strategies used in past offset events and comparing their compensation effects. Each type of thermal displacement mismatch root cause corresponds to a set of optimal mismatch compensation strategies.

[0028] Obtain the sequential root cause list. For the first type of thermal displacement mismatch root cause in the sequential root cause list, find the corresponding optimal mismatch compensation strategy in the compensation strategy library and execute it as the current mismatch compensation strategy. Set an execution buffer period starting from the time of executing the mismatch compensation strategy. At the end of the execution buffer period, deduce and calculate the actual thermal displacement prediction vector of the wafer test and determine whether there is a risk of thermal displacement mismatch. If there is, obtain the next type of thermal displacement mismatch root cause in the sequential root cause list and compensate for it until there is no risk of thermal displacement mismatch at the end of the corresponding execution buffer period, or there is no next type of thermal displacement mismatch root cause in the sequential root cause list.

[0029] A wafer testing offset detection system includes the following modules:

[0030] Acquisition module: Acquires key vector data during wafer testing, including thermal load vector, system state vector, and real-time observation vector, and organizes them to obtain a multimodal dataset;

[0031] Risk prediction module: Establish a hybrid prediction model that combines physical simulation model and residual correction model. Based on multimodal dataset, use the hybrid prediction model to predict the actual thermal displacement prediction vector, determine whether there is a risk of thermal displacement mismatch, and if so, trigger offset root cause identification.

[0032] Root cause matching module: Establishes a historical knowledge base. If offset root cause identification is triggered, it identifies the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generates a sequential root cause list.

[0033] Risk compensation module: Based on the obtained sequential root cause list, generate mismatch compensation strategies for thermal displacement mismatch root causes in sequence to compensate for thermal displacement mismatch risks.

[0034] The beneficial effects of this invention are as follows:

[0035] 1. This invention collects key vector data during the testing process and organizes it into a multimodal dataset, providing a comprehensive and accurate foundation for subsequent analysis. By establishing a hybrid prediction model, it can predict the risk of thermal displacement mismatch in advance, trigger the identification of the root cause of the offset in a timely manner, solve the problem in its infancy, avoid wafer test failure due to thermal displacement mismatch, effectively reduce production costs, improve production efficiency, and ensure the stability of product quality.

[0036] 2. This invention establishes a historical knowledge base and combines it with a multimodal dataset to quickly and accurately identify the root causes of thermal displacement mismatch and generate a sequential root cause list. Based on this list, a mismatch compensation strategy is generated in sequence, which can effectively compensate for the risk of thermal displacement mismatch, improve the pertinence and effectiveness of compensation, reduce the waste of resources caused by blind adjustments, and further improve the accuracy and reliability of wafer testing. Attached Figure Description

[0037] The invention will now be further described with reference to the accompanying drawings.

[0038] Figure 1 This is a flowchart illustrating the steps of a wafer testing offset detection method according to an embodiment of the present invention;

[0039] Figure 2 This is a module architecture diagram of a wafer testing offset detection system according to an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of a wafer test offset detection storage medium according to an embodiment of the present invention. Detailed Implementation

[0041] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 As shown in the figure, a wafer testing offset detection method according to an embodiment of the present invention includes the following steps:

[0044] S1: During the wafer testing process, key vector data is collected, including thermal load vector, system state vector and real-time observation vector, and then processed to obtain a multimodal dataset.

[0045] Specifically, thermal load vector Used to characterize changes in the external thermal environment, the components of the thermal load vector include the ambient temperature sequence Tenv(t) and the power consumption spectrum Ptest(i,t). The ambient temperature sequence Tenv(t) is acquired by a platinum resistance sensor built into the wafer test cavity, where t represents time. It records the change curve of the cavity ambient temperature over time during the wafer test in real time. The power consumption spectrum Ptest(i,t) is acquired by the power management module of the wafer test equipment. Ptest(i,t) represents the real-time power consumption value of the i-th test item at time t. The power consumption spectrum reflects the contribution of the self-heating of the test item to the local thermal displacement during the wafer test.

[0046] System state vector Used to reflect the state of the test system itself, the components of the system state vector include (λ, E, α), the device runtime trun, and the device temperature Tdev. Among them, in the probe card material parameters, λ represents the thermal conductivity, E represents the elastic modulus, and α represents the thermal expansion coefficient. The probe card material parameters are obtained from the probe card manufacturer's material manual. The device runtime trun is obtained by accumulating time from the device control system log. The device temperature Tdev is measured by the temperature sensor set in the core component of the device, reflecting the impact of the device's self-heating on the test accuracy.

[0047] Real-time observation vector To capture subtle changes, the components of the real-time observation vector include a high-frequency visual marker image Imark(t) and a contact resistance micro-change ΔRcon(t). The high-frequency visual marker image Imark(t) is captured by a wafer surface marking camera at a preset frame rate and is used to capture subtle changes in the wafer position. The contact resistance micro-change ΔRcon(t) is measured by a probe wafer contact resistance detection module. Since the contact resistance is positively correlated with the probe position, the contact resistance micro-change reflects the probe position offset.

[0048] Using the time t of the high-frequency visually labeled images as the reference time, linear interpolation is used to align the components of the thermal load vector, system state vector, and real-time observation vector that change with time t, ensuring that all key vector data match in the same time dimension, thus obtaining a time-synchronized multimodal dataset:

[0049] ;

[0050] Where t represents the baseline time, and the interval between all baseline times is the same; T represents the total number of baseline times in the multimodal dataset.

[0051] It should be noted that the purpose of this step is to systematically collect multi-dimensional and heterogeneous data from different sources during the wafer testing process and synchronize them in time to form a unified and aligned multimodal dataset. This dataset covers internal and external factors that affect test offsets, as well as direct and indirect observational evidence, providing a solid data foundation for accurate analysis. By using high-frequency vision as a benchmark for linear interpolation synchronization, it ensures that all data are comparable on the same time axis, avoiding timing misalignment and analysis errors caused by different data acquisition frequencies.

[0052] S2: Establish a hybrid prediction model that combines a physical simulation model and a residual correction model. Based on a multimodal dataset, use the hybrid prediction model to predict the actual thermal displacement prediction vector and determine whether there is a risk of thermal displacement mismatch. If so, trigger the identification of the root cause of the offset.

[0053] Specifically, the physical simulation model is based on the thermo-structure coupling theory and calculates the displacement of the wafer and probe caused by the Caine thermal effect from the perspective of physical principles, including a temperature field solution sub-model and a structural displacement calculation sub-model.

[0054] For the temperature field solution sub-model, obtain the thermal load vector from the multimodal dataset. As input, the temperature fields of the wafer and probe card are solved based on the heat conduction equation. The heat conduction equation is:

[0055] ;

[0056] Where x, y, and z represent three mutually perpendicular coordinate axes in three-dimensional space, used to determine the spatial position of any point on the wafer and probe card, and t represents time. The density of the material is indicated by a probe card and obtained from the material datasheet of the wafer; c represents the specific heat capacity of the material. Indicates the thermal conductivity coefficient. The intensity of the internal heat source is represented by the heat load vector. The power consumption spectrum Ptest(i,t) of the test item is obtained by combining the volume of the heat-generating area;

[0057] It should be noted that the role of internal heat source intensity is to reflect the power consumption differences of different test items in wafer testing into the temperature field calculation. For example, when a high-power test item is running, its corresponding internal heat source intensity... The value will increase, leading to a rise in local temperature;

[0058] Based on thermal load vector The boundary conditions of the heat conduction equation are determined by the ambient temperature sequence Tenv(t). By solving the heat conduction equation, the temperature fields of the wafer and probe card as a function of time and space are obtained. ;

[0059] For the structural displacement calculation sub-model, based on Hooke's law and combined with the calculated temperature field... Calculate the theoretical thermal displacements in the x and y directions caused by thermal expansion. and The calculation formula is:

[0060] ;

[0061] ;

[0062] in, Indicates the coefficient of thermal expansion of a material. This indicates the ambient temperature of the cavity at the start of wafer testing. The wafer surface height is indicated by calibration data from the testing equipment. This represents the temperature change obtained based on the temperature field. and These represent the start and end positions of integration in the x-direction, determined based on the dimensions of the wafer and probe card in the x-direction. and These represent the start and end positions of integration in the y-direction, respectively, and are determined based on the size range of the wafer and probe card in the y-direction.

[0063] The theoretical thermal displacements obtained in the x and y directions and By integrating the data, the theoretical thermal displacement vector is obtained. ;

[0064] It should be noted that physical simulation models can calculate thermal displacement vectors based on physical principles, but in actual wafer testing environments, there are many nonlinear factors that cannot be described by accurate physical modeling, such as probe wear, airflow disturbance, and minor equipment vibration. Therefore, the hybrid prediction model introduces a residual correction model to make up for the limitations of physical simulation models.

[0065] The residual correction model uses a spatiotemporal graph neural network (ST-GNN) to learn the error between the theoretical thermal displacement vector predicted by the physical simulation model and the actual displacement vector, and corrects the physical simulation model accordingly.

[0066] Specifically, a multimodal dataset is obtained, and Z-score standardization is performed on all key vector data in the multimodal dataset to unify the value range of all key vector data. An attention mechanism fusion algorithm is used to assign weights to the standardized thermal load vector, system state vector and real-time observation vector in the multimodal dataset and perform weighted fusion to obtain a unified multimodal feature matrix.

[0067] The multimodal feature matrix and the theoretical thermal displacement vector are used as inputs to the residual correction model. The network structure of the residual correction model consists of three spatiotemporal convolutional layers and one fully connected output layer. The spatiotemporal convolutional layers are used to capture the correlation between time series and spatial location, and output the predicted residual vector at time t. ;

[0068] A historical time period is defined, and the theoretical thermal displacement vector and the corresponding actual thermal displacement vector obtained through detection within this period are acquired. The actual residual vector within the historical time period is calculated. The multimodal feature matrix, theoretical thermal displacement vector, and actual residual vector within the historical time period are integrated to obtain a corrected training set. The corrected training set is used to train the residual correction model, and the trained residual correction model is used for derivation to obtain the predicted residual vector. And output;

[0069] The theoretical displacement vector calculated from the physical simulation model The predicted residual vector of the residual correction model By superimposing the vectors, the actual thermal displacement prediction vector is obtained. :

[0070] ;

[0071] To evaluate the actual thermal displacement prediction vector To ensure reliability, a confidence assessment mechanism based on Bayesian networks is introduced to calculate the prediction confidence. If the prediction confidence is less than the preset confidence standard, the prediction is judged to be inaccurate and manual verification is triggered. Otherwise, the prediction is judged to be accurate. The actual thermal displacement prediction vector is compared with the thermal displacement vector threshold. If the actual thermal displacement prediction vector exceeds the thermal displacement vector threshold, it is judged that there is a risk of thermal displacement mismatch, the wafer test has shifted, and the root cause of the shift is triggered. Otherwise, the wafer test is judged to be normal.

[0072] It should be noted that this step establishes a hybrid prediction model that combines a physical simulation model and a residual correction model. This model is used to predict the risk of thermal displacement mismatch in wafer testing with high accuracy and confidence. The physical simulation model provides an interpretable theoretical basis that conforms to physical laws, while the residual correction model learns the errors caused by nonlinear perturbations that the physical model cannot cover. The combination of the two achieves accurate predictions based on first principles and empirical corrections. The introduction of Bayesian confidence assessment not only provides the predicted value but also the degree of confidence of the predicted value. When the confidence is low, manual verification is triggered, avoiding the risks of blindly trusting the model, improving the robustness and security of the system, and realizing an automated process from data to risk assessment, replacing the method of relying on manual experience. The combination of the thermal-structure coupled physical simulation model and the residual correction model based on spatiotemporal graph neural network is an innovative application of physical information and artificial intelligence in the field of wafer testing. It combines the interpretability of the physical simulation model with the advantages of spatiotemporal graph neural network in handling complex nonlinearities.

[0073] S3: Establish a historical knowledge base. If offset root cause identification is triggered, identify the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generate a sequential root cause list.

[0074] Specifically, the historical knowledge base stores feature-cause mapping pairs of past offset events. Past offset events represent events with thermal displacement mismatch risk during historical periods. The feature-cause mapping pairs include the multimodal dataset of past offset events, the corresponding actual thermal displacement vectors, and the root causes of thermal displacement mismatch.

[0075] If offset root cause identification is triggered, a multimodal dataset is obtained, and the cosine similarity between the multimodal dataset and the multimodal dataset of past offset events stored in the historical knowledge base is calculated. If the cosine similarity is greater than the preset similarity standard, it is determined that the past offset event matches the current one, and the hot offset mismatch root cause of the past offset event is included in the candidate root cause list.

[0076] If there is no thermal displacement mismatch root cause in the candidate root cause list, a risk warning is sent to the administrator terminal and the thermal displacement mismatch risk is manually handled. If there is only one thermal displacement mismatch root cause in the candidate root cause list, the thermal displacement mismatch root cause is determined to be the current thermal displacement mismatch root cause, and a sequential root cause list is generated. The sequential root cause list contains only one thermal displacement mismatch root cause.

[0077] If there are multiple types of thermal displacement mismatch root causes in the candidate root cause list, prioritize the thermal displacement mismatch root causes in the candidate root cause list to generate an ordered root cause list.

[0078] Specifically, the number of times the thermal displacement mismatch root cause appears in the candidate root cause list is obtained, and the ratio is calculated with the total number of times the thermal displacement mismatch root cause appears in the candidate root cause list to obtain the frequency of thermal displacement mismatch root cause in the candidate root cause list.

[0079] For any type of thermal displacement mismatch root cause in the candidate root cause list, obtain all actual thermal displacement vectors corresponding to the thermal displacement mismatch root cause in the historical knowledge base and perform average processing to obtain the average thermal displacement vector of the thermal displacement mismatch root cause. Normalize the average thermal displacement vector of all thermal displacement mismatch root causes to obtain the normalized thermal displacement vector of each type of thermal displacement mismatch root cause.

[0080] The Analytic Hierarchy Process (AHP) is used to assign weights to the frequency of occurrence of thermal displacement mismatch root causes in the candidate root cause list and the normalized thermal displacement vector. Weighted fusion processing is then performed to obtain the priority coefficients of thermal displacement mismatch root causes. Based on the priority coefficients, the thermal displacement mismatch root causes in the candidate root cause list are sorted in descending order to generate the sequential root cause list.

[0081] It should be noted that the purpose of this step is to automatically identify the most likely root cause of the deviation after detecting the risk of deviation by comparing it with the historical case library, and to prioritize them. By using historical experience and similarity matching, candidate root causes can be quickly identified, which greatly shortens the troubleshooting time and reduces production interruptions. When there are multiple candidate causes, intelligent sorting is performed based on frequency and severity analytic hierarchy process to generate repair order suggestions, guiding engineers to handle them according to priority and improving troubleshooting efficiency.

[0082] S4: Based on the obtained sequential root cause list, generate mismatch compensation strategies for thermal displacement mismatch root causes in sequence to compensate for the risk of thermal displacement mismatch.

[0083] Specifically, a compensation strategy library is established, which stores the optimal mismatch compensation strategy corresponding to various thermal displacement mismatch root causes. The strategy is selected by obtaining the mismatch compensation strategies used in past offset events and comparing their compensation effects. The compensation strategy library adopts a structured storage method, with a set of optimal mismatch compensation strategies corresponding to each type of thermal displacement mismatch root cause.

[0084] For example, if the thermal displacement mismatch is caused by a sudden change in ambient temperature, the optimal mismatch compensation strategy is to adopt a global compensation strategy of a precision motion platform. The compensation amount is calculated based on the actual thermal displacement prediction vector. The compensation amounts in the x and y directions are opposite to the actual thermal displacement prediction vector and equal in magnitude. The ambient temperature of the cavity is adjusted according to the preset temperature adjustment rate and the target temperature to reduce the impact of ambient temperature on wafer testing.

[0085] Obtain the sequential root cause list. For the first type of thermal displacement mismatch root cause in the sequential root cause list, obtain the optimal mismatch compensation strategy corresponding to the thermal displacement mismatch root cause stored in the compensation strategy library. Use the optimal mismatch compensation strategy as the current mismatch compensation strategy, calculate the compensation amount and execute the mismatch compensation strategy. Set an execution buffer period starting from the time of execution of the mismatch compensation strategy. At the end of the execution buffer period, deduce and calculate the actual thermal displacement prediction vector of the wafer test and determine whether there is a risk of thermal displacement mismatch. If there is, and there is still a next type of thermal displacement mismatch root cause in the sequential root cause list, obtain the optimal mismatch compensation strategy corresponding to the next type of thermal displacement mismatch root cause for compensation, until there is no risk of thermal displacement mismatch at the end of the corresponding execution buffer period, or there is no next type of thermal displacement mismatch root cause in the sequential root cause list.

[0086] If there is no next type of thermal displacement mismatch root cause in the sequential root cause list, there is still a risk of thermal displacement mismatch at the end of the corresponding execution buffer period. A risk warning is sent to the administrator terminal, and the thermal displacement mismatch risk is handled manually.

[0087] It should be noted that the purpose of this step is to automatically call the preset optimal compensation strategy from the strategy library based on the identified root cause and execute it, forming a closed-loop control loop of detection-identification-compensation. This transforms the traditional problem-finding shutdown and troubleshooting into proactive compensation for predicted problems, enabling intervention before test results are affected. This effectively improves equipment utilization and production cycle time. The strategies in the compensation strategy library are derived from historical best practices and have different strategies for different root causes, achieving targeted solutions and higher compensation efficiency. The compensation strategy is tried in the order of the generated root cause list, and the effect is verified after each attempt. The design logic is rigorous and can effectively cope with complex offset scenarios caused by multiple concurrent factors.

[0088] The technical solution of this invention is as follows: During wafer testing, key vector data is collected, including thermal load vector, system state vector, and real-time observation vector. A multimodal dataset is obtained, and a hybrid prediction model combining a physical simulation model and a residual correction model is established. Based on the multimodal dataset, the hybrid prediction model is used to predict the actual thermal displacement prediction vector. It is determined whether there is a risk of thermal displacement mismatch. If so, the root cause identification of the offset is triggered, and a historical knowledge base is established. If the root cause identification of the offset is triggered, the root cause of thermal displacement mismatch in wafer testing is identified based on the multimodal dataset and the historical knowledge base, and a sequential root cause list is generated. According to the obtained sequential root cause list, a mismatch compensation strategy is generated sequentially for the root cause of thermal displacement mismatch to compensate for the risk of thermal displacement mismatch.

[0089] Example 2:

[0090] like Figure 2 As shown in the figure, a wafer testing offset detection system according to an embodiment of the present invention includes the following modules:

[0091] Acquisition module: Acquires key vector data during wafer testing, including thermal load vector, system state vector, and real-time observation vector, and organizes them to obtain a multimodal dataset;

[0092] Risk prediction module: Establish a hybrid prediction model that combines physical simulation model and residual correction model. Based on multimodal dataset, use the hybrid prediction model to predict the actual thermal displacement prediction vector, determine whether there is a risk of thermal displacement mismatch, and if so, trigger offset root cause identification.

[0093] Root cause matching module: Establishes a historical knowledge base. If offset root cause identification is triggered, it identifies the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generates a sequential root cause list.

[0094] Risk compensation module: Based on the obtained sequential root cause list, generate mismatch compensation strategies for thermal displacement mismatch root causes in sequence to compensate for thermal displacement mismatch risks.

[0095] Example 3:

[0096] Reference Figure 3 The present invention also provides a wafer test offset detection storage medium, namely a computer 3, comprising: a memory 302 and a processor 301 and a computer program 303 stored on the memory 302. When the computer program 303 is executed on the processor 301, a wafer test offset detection method as described in any of the above methods is implemented.

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A wafer testing offset detection method, characterized in that: include: Key vector data is collected during wafer testing. The key vector data includes thermal load vector, system state vector and real-time observation vector, which are then processed to obtain a multimodal dataset. A hybrid prediction model combining a physical simulation model and a residual correction model is established. Based on a multimodal dataset, the hybrid prediction model is used to predict the actual thermal displacement prediction vector. It is then used to determine whether there is a risk of thermal displacement mismatch. If so, the root cause identification of the offset is triggered. The physical simulation model includes a temperature field solution sub-model and a structural displacement calculation sub-model; The temperature field solution sub-model and the structural displacement calculation sub-model include: In the temperature field solution sub-model, the thermal load vector in the multimodal dataset is obtained as input. The thermal load vector includes the ambient temperature sequence and the power consumption spectrum of the test item. The boundary conditions of the heat conduction equation are determined according to the ambient temperature sequence. The internal heat source intensity is calculated by combining the power consumption spectrum of the test item and the volume of the heat-generating area. Based on the internal heat source intensity, the temperature field of the wafer and probe card that changes with time and space is obtained by solving the heat conduction equation and output. In the structural displacement calculation sub-model, the temperature field output by the temperature field solution sub-model is used as input. Based on Hooke's law, the theoretical thermal displacements in the x and y directions caused by thermal expansion are calculated and integrated to obtain the theoretical thermal displacement vector. The residual correction model includes: A multimodal dataset is obtained. All key vector data in the multimodal dataset are Z-score normalized and fused using an attention mechanism fusion algorithm to obtain a multimodal feature matrix. The multimodal feature matrix and the theoretical thermal displacement vector are used as inputs to the residual correction model. The network structure of the residual correction model includes three spatiotemporal convolutional layers and one fully connected output layer, which outputs the predicted residual vector. The actual thermal displacement prediction vector is obtained as follows: Set a historical time period, obtain the theoretical thermal displacement vector within the historical time period and the corresponding actual thermal displacement vector obtained through detection and acquisition, calculate the corresponding actual residual vector, integrate it with the multimodal feature matrix to obtain the correction training set, train the residual correction model, use the trained residual correction model to perform deduction to obtain the predicted residual vector, and superimpose the theoretical displacement vector calculated by the physical simulation model with the predicted residual vector predicted by the residual correction model to obtain the actual thermal displacement prediction vector. Establish a historical knowledge base. If offset root cause identification is triggered, identify the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generate a sequential root cause list. The historical knowledge base is established in the following way: Establish a historical knowledge base to store feature-cause mapping pairs of past offset events, set historical time periods, and past offset events represent events with thermal displacement mismatch risk within the historical time period. The feature-cause mapping pairs include the multimodal dataset of past offset events, the corresponding actual thermal displacement vectors collected, and the root causes of thermal displacement mismatch. The method for identifying the root cause of thermal displacement mismatch is as follows: If offset root cause identification is triggered, a multimodal dataset is obtained, and the cosine similarity between the multimodal dataset and the multimodal dataset of past offset events stored in the historical knowledge base is calculated. If the cosine similarity is greater than the preset similarity standard, the thermal displacement mismatch root cause of the past offset event is included in the candidate root cause list as the identified thermal displacement mismatch root cause. The sequential root cause list is generated as follows: The frequency of occurrence of thermal displacement mismatch root causes in the candidate root cause list is obtained and processed with the total number of occurrences of thermal displacement mismatch root causes in the candidate root cause list to obtain the frequency of occurrence of thermal displacement mismatch root causes. All actual thermal displacement vectors corresponding to thermal displacement mismatch root causes in the historical knowledge base are obtained and processed to obtain the normalized thermal displacement vector of each type of thermal displacement mismatch root cause. The analytic hierarchy process (AHP) is used to assign weights to the frequency of occurrence of thermal displacement mismatch root causes and the normalized thermal displacement vector and perform weighted fusion processing to obtain the priority coefficient of thermal displacement mismatch root causes. The thermal displacement mismatch root causes in the candidate root cause list are sorted in descending order based on the priority coefficient to generate the ordered root cause list. Based on the obtained sequential root cause list, mismatch compensation strategies are generated sequentially for thermal displacement mismatch root causes to compensate for the risk of thermal displacement mismatch.

2. The wafer test offset detection method according to claim 1, characterized in that: The method for compensating for the risk of thermal displacement mismatch is as follows: A compensation strategy library is established, which stores the optimal mismatch compensation strategy corresponding to various thermal displacement mismatch root causes. The optimal mismatch compensation strategy is selected by obtaining the mismatch compensation strategies used in past offset events and comparing their compensation effects. Each type of thermal displacement mismatch root cause corresponds to a set of optimal mismatch compensation strategies. Obtain the sequential root cause list. For the first type of thermal displacement mismatch root cause in the sequential root cause list, find the corresponding optimal mismatch compensation strategy in the compensation strategy library and execute it as the current mismatch compensation strategy. Set an execution buffer period starting from the time of executing the mismatch compensation strategy. At the end of the execution buffer period, deduce and calculate the actual thermal displacement prediction vector of the wafer test and determine whether there is a risk of thermal displacement mismatch. If there is, obtain the next type of thermal displacement mismatch root cause in the sequential root cause list and compensate for it until there is no risk of thermal displacement mismatch at the end of the corresponding execution buffer period, or there is no next type of thermal displacement mismatch root cause in the sequential root cause list.

3. A wafer testing offset detection system, used to implement the wafer testing offset detection method according to any one of claims 1-2, characterized in that: Includes the following modules: Acquisition module: Acquires key vector data during wafer testing, including thermal load vector, system state vector, and real-time observation vector, and organizes them into a multimodal dataset; Risk prediction module: Establish a hybrid prediction model that combines physical simulation model and residual correction model. Based on multimodal dataset, use the hybrid prediction model to predict the actual thermal displacement prediction vector, determine whether there is a risk of thermal displacement mismatch, and if so, trigger offset root cause identification. Root cause matching module: Establishes a historical knowledge base. If offset root cause identification is triggered, it identifies the thermal displacement mismatch root cause of wafer testing based on the multimodal dataset and the historical knowledge base, and generates a sequential root cause list. Risk compensation module: Based on the obtained sequential root cause list, generate mismatch compensation strategies for thermal displacement mismatch root causes in sequence to compensate for thermal displacement mismatch risks.

4. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed on a processor, implements the steps of the wafer test offset detection method according to any one of claims 1-2.

Images